Research Engineer, World Models

We are looking for a Research Engineer in World Models. In this role, you will build multi-modal generative models (video, audio, text, actions) that predict all of NEO’s future sensor readings and actions from its past observations and actions. Such foundation models enable general understanding of unstructured environments.

Responsibilities:

• Full-stack engineering, from data engineering to model architecture design to shipping polished products
• Translate improvements in world models to improvements in robot autonomy.
• Implement high-throughput data loaders for large multi-modal datasets.
• Implement tokenizers and transformers to fit web-scale robot data.
• Predict how real-world robot performance scales with pre-training metrics (e.g. log loss).

Must-Haves:

• Getting general-purpose robots to work in the home is just about the hardest problem one can work on. We are looking for people with the courage to tackle unsolved technical challenges with an intense work ethic.
• 4+ years of experience of Python and Deep Learning Frameworks (TensorFlow, PyTorch, JAX, Keras).
• Experience training a variety of deep generative models on video.
• Ability to rapidly iterate on ideas and scale up promising ones to increasing amounts of compute and data.
• In your cover letter, please tell us about one idea in world modeling that should work but doesn’t work quite yet, and why you think that’s the case.

Nice-to-Haves:

• Strong background in distributed systems and parallel computing.
• Experience in large-scale model training (>70B parameters)
• Experience with the 1X World Modeling Challenge (Compression, Sampling, Evaluation).
• Experience writing testing infrastructure for data loaders and robotics datasets.

Sample Projects

We encourage you to apply even if you do not meet every single qualification. If you think you can do an exceptional job at one of the “sample projects” listed below, please let us know in your cover letter.

• Converting World Models into a policy via visual MPC, and making sampling fast enough to do dynamic motions.
• Developing evaluation metrics for policies simulated inside world models (e.g. task success prediction from a generated video).
• Using the World Model for tasks beyond evaluation and training.

Compensation
At 1X your work and results will be rewarded with a total rewards package consisting of a base salary, stock options and benefits. Base salary range is $130,000 to $250,000. Your actual salary will be based on your knowledge, skills and experience.

Location Policy

We believe the best work is done when collaborating and therefore require in-person presence in our office locations.

We're excited to get to know you and the prospects of having you on board!